Sampling strategies in ordinal regression for surrogate assisted evolutionary optimization

In evolutionary optimization surrogate models are commonly used when the evaluation of a fitness function is computationally expensive. Here the fitness of individuals are indirectly estimated by modeling their rank with respect to the current population by use of ordinal regression. This paper focuses on how to validate the goodness of fit for surrogate models during search and introduces a novel validation/updating policy for surrogate models, and is illustrated on classical numerical optimization functions for evolutionary computation. The study shows that for validation accuracy it is sufficient for the approximate ranking and true ranking of the training set to be sufficiently concordant or that only the potential parent individuals should be ranked consistently. Moreover, the new validation approach reduces the number of fitness evaluation needed, without a loss in performance.